Exam technique

Never the answer on the Claude Architect exam

Some Claude Certified Architect (CCAR-F) options are wrong in every scenario. This lists the recurring ones and the truth that replaces each.

SpacedPublished 10 Oct 20263 min read

Some CCAR-F options are wrong only in context, right for a different ask or a different scenario, or right but not first. Others are wrong everywhere, because each one misstates how the mechanism works. When one of these appears you can cross it out without reading the rest of the question.

Agents and loops

The claim	The truth
Raise the iteration cap to fix an agent that stops early	Read `stop_reason`: the loop continues on `tool_use` and ends on `end_turn`.
Text in the response means the agent has finished	Text arrives alongside `tool_use` blocks. Only `stop_reason` tells talking from finishing.
Iteration caps are the main way to control a loop	`stop_reason` is the control. A cap is a safety net.
Subagents can read the coordinator's context	Each starts empty. Everything it needs is passed explicitly.
Subagents can message each other directly	Everything routes through the coordinator.
A rule in the system prompt guarantees compliance	Prompts are probabilistic. A hook enforces a rule every time.
Hooks are prompt-based guardrails	Hooks are deterministic code.

Rule · remember thisPrompt versus codeIf a single slip is unacceptable, the answer is code that runs before the action.

Tools and MCP

The claim	The truth
The tool name drives which tool the model picks	The description does.
Cut the number of tools to fix misselection	Improve the descriptions first. Split or consolidate only past a handful of tools per agent.
One agent with every tool is simpler	Focused agents with a few tools each choose better.
`tool_choice: any` forces a particular tool	It guarantees only that some tool is called. Name the tool to force it.
Empty results mean the tool failed	No data is a valid result. A query that never ran is the error.
Return a plain error string	Return structured error metadata the model can act on.
MCP servers connect to each other	They connect through the client and the host, never to each other.
Build a custom server for a common integration	Check for a community server first.

Claude Code

The claim	The truth
A user-level CLAUDE.md overrides the project's	The files are combined. None overrides another.
Put every rule in the project CLAUDE.md	Use `.claude/rules/` for rules that only apply to certain paths.
A single `.md` file in `.claude/skills/` makes a skill	A skill is a directory with a `SKILL.md`. A flat file is a command.
`-p` turns on plan mode	`-p` runs Claude Code non-interactively. Plan mode is for ambiguous tasks.
Review code in the session that wrote it	Review it in an independent session that shares none of that context.
`--append-system-prompt` replaces the system prompt	It adds to it.

Prompts, output and context

The claim	The truth
More few-shot examples are always better	A few well-chosen examples, including an edge case, beat a long list.
A required field stops the model making things up	A required field with no data forces invention. Make it nullable.
JSON asked for in the prompt is fine in production	Force a tool call when the schema has to hold.
Say try again after a validation failure	Resend the original prompt with the failed output and the exact validation error.
The Batch API is faster	It is cheaper and slower. Use it only when nobody is waiting.
Summarise as you go to save context	Carry the specific facts forward word for word in a persistent block.
A larger context window fixes lost details in a long conversation	Capacity is almost never the fix.
Keep everything in the conversation history	Write findings to a scratchpad file. What is on disk cannot degrade.
A healthy overall error rate means nothing is wrong	Silent failures hide inside a good average. Surface errors per type.
High aggregate accuracy means the output is reliable	Measure by document type and field; one strong category can hide a weak one.
Pick the more recent source when two disagree	Keep both and annotate each with its date.

The post on Claude Certified Architect wrong-answer patterns covers how to spot the context-dependent options by their shape.